Teaching large language models (LLMs) to be faithful in the provided context is crucial for building reliable information-seeking systems. Therefore, we propose a systematic framework, CANOE, to reduce faithfulness hallucinations of LLMs across different downstream tasks without human annotations. Specifically, we first synthesize short-form question-answering (QA) data with four diverse tasks to construct high-quality and easily verifiable training data without human annotation. Also, we propose Dual-GRPO, a rule-based reinforcement learning method that includes three tailored rule-based rewards derived from synthesized short-form QA data, while simultaneously optimizing both short-form and long-form response generation. Notably, Dual-GRPO eliminates the need to manually label preference data to train reward models and avoids over-optimizing short-form generation when relying only on the synthesized short-form QA data. Experimental results show that CANOE greatly improves the faithfulness of LLMs across 11 different tasks, even outperforming the most advanced LLMs, e.g., GPT-4o and OpenAI o1.

Recent advancements in large language models (LLMs) have led to significant breakthroughs in mathematical reasoning capabilities. However, existing benchmarks like GSM8K or MATH are now being solved with high accuracy (e.g., OpenAI o1 achieves 94.8% on MATH dataset), indicating their inadequacy for truly challenging these models. To bridge this gap, we propose a comprehensive and challenging benchmark specifically designed to assess LLMs' mathematical reasoning at the Olympiad level. Unlike existing Olympiad-related benchmarks, our dataset focuses exclusively on mathematics and comprises a vast collection of 4428 competition-level problems with rigorous human annotation. These problems are meticulously categorized into over 33 sub-domains and span more than 10 distinct difficulty levels, enabling a holistic assessment of model performance in Olympiad-mathematical reasoning. Furthermore, we conducted an in-depth analysis based on this benchmark. Our experimental results show that even the most advanced models, OpenAI o1-mini and OpenAI o1-preview, struggle with highly challenging Olympiad-level problems, with 60.54% and 52.55% accuracy, highlighting significant challenges in Olympiad-level mathematical reasoning.

Fine-grained geographic named entity recognition involves the identification of precise locations within geographic text. Existing methods face challenges in effectively addressing this task, namely: (1) the complexity and long-dependency nature of various fine-grained locations, which pose difficulties for recognition, and (2) the presence of numerous out-of-vocabulary fine-grained locations that are not present in the training data . In this paper, we propose a Multi-scale Siamese Reconstruction Network (MSRN) to tackle these challenges. To address the first challenge, MSRN employs a multi-scale convolutional network to aggregate interactions among coarse-grained and fine-grained span features. To tackle the second challenge, MSRN introduces a siamese reconstruction network that corrupts the representation and then reconstructs it, preventing the model from relying on rote memorization of biased locations in the training data . This approach enhances the model’s understanding of entity context. We conduct experiments on four real-world fine-grained geographic named entity recognition datasets, which contain a variety of complex long fine-grained locations and geographic entities not present in the training data. These experiments illustrate the challenges commonly encountered in real-world engineering applications and deliver a solid foundation for evaluating the robustness and applicability of our proposed method. The experimental results demonstrate the superiority of MSRN over previous models, achieving significant performance improvements ranging from 2.1 to 3.0. Further analysis confirms the effectiveness of MSRN in recognizing complex long-dependency entities, with a performance improvement of 3.0, as well as out-of-vocabulary locations, with a performance improvement of 6.6 compared to models with similar parameter sizes.